Optimized Low-VRAM Multi-LoRA Workflow for Flux.1 Dev (GGUF)

Another new, refined and enhanced workflow by Experimental_dAIver

This workflow delivers the full power of Flux.1 Dev in GGUF format, specially optimized for GPUs with less than 8 GB VRAM, like my RTX 4050 with only 6 GB. The integrated CacheDiT_Model_Optimizer and PatchSageAttentionKJ (both optional!) provide a noticeable speed boost with almost no quality loss. Full NdSuperLoraLoader support for easy multi-LoRA stacking with automatic trigger-word integration, the intuitive selectLatentSizePlus aspect-ratio and resolution selector, plus a complete SEEDVR2 Video Upscaler Subgraph (base provided by @LumaRift) for stunning high-resolution results complete this elegant and highly flexible setup.

Version 1.5 is my first public release and brings significant improvements in speed, usability, multi-LoRA handling and upscaling quality — while remaining extremely VRAM-efficient (tested on RTX 4050 with only 6 GB).

Key features in v1.5:

• NdSuperLoraLoader with full multi-LoRA support, auto-fetch trigger words and easy strength control

• selectLatentSizePlus — intuitive aspect-ratio and resolution selector with beautiful presets (including golden-ratio-friendly options) plus easy orientation swap

• Full SEEDVR2 Video Upscaler Subgraph — powerful DiT-based high-end upscaler that delivers stunning 4K+ results with intelligent resolution handling, Lab color correction and temporal settings. Works exceptionally well on still images too, producing superior detail and coherence

• Improved workflow organization, expanded notes with recommended settings, and more robust saving options via SaveImageExtended

• CacheDiT and SageAttention integration for optimized Flux performance on low-VRAM hardware

Required Custom Nodes

• ComfyUI-GGUF - https://github.com/city96/ComfyUI-GGUF - UnetLoaderGGUF + CLIPLoaderGGUF

• nd-super-nodes - https://github.com/HenkDz/nd-super-nodes - NdSuperLoraLoader with tags, trigger words & beautiful UI

• save-image-extended-comfyui - https://github.com/thedyze/save-image-extended-comfyui - Advanced saving with metadata & dynamic filenames

• ComfyUi-MzMaXaM - https://github.com/MzMaXaM/ComfyUi-MzMaXaM - selectLatentSizePlus

• ComfyUI-SeedVR2_VideoUpscaler - https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler - SEEDVR2 Video Upscaler Subgraph

Models & Downloads (exact paths)

The following list explains the base models I am most frequently using with this workflow. The list as well explains where to put each file after you downloaded it.

1. Main Model (Flux.1 Dev GGUF):

• File: flux1-dev-Q5_1.gguf (or any other GGUF quant you prefer – Q4, Q5, Q6 or Q8 all work well)

• Download: https://huggingface.co/city96/FLUX.1-dev-gguf/resolve/main/flux1-dev-Q5_1.gguf

• Target folder: ComfyUI/models/diffusion_models/ (or Flux/ subfolder)

2. Text Encoder (CLIP)

• File: t5xxl_fp8_e4m3fn.safetensors + clip_l.safetensors (standard Flux pair)

• Download t5xxl: https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/t5xxl_fp8_e4m3fn.safetensors
  Download clip_l: https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/clip_l.safetensors

• Target folder: ComfyUI/models/clip/ or ComfyUI/models/text_encoders/

3. VAE

• File: ae.safetensors

• Download: https://huggingface.co/Comfy-Org/flux1-dev/resolve/main/split_files/vae/ae.safetensors

• Target folder: ComfyUI/models/vae/

4. Upscalers

• 4× Upscaler: 4x-UltraSharp → https://civarchive.com/models/116225/4x-ultrasharp

• 1× Skin-Contrast Upscaler: 1xSkinContrast-High-SuperUltraCompact.pth → https://huggingface.co/notkenski/upscalers/blob/main/1xSkinContrast-High-SuperUltraCompact.pth

• Target folder: ComfyUI/models/upscale_models/

5. LoRAs (for NdSuperLoraLoader – full multi-LoRA support)

• Any Flux-compatible LoRA (.safetensors)

• Target folder: ComfyUI/models/loras/

6. SEEDVR2 Models (for the high-end upscaler subgraph – optional but recommended)

• DiT Model: seedvr2_ema_3b-Q8_0.gguf

• VAE: ema_vae_fp16.safetensors

• Download from the official ComfyUI-SeedVR2_VideoUpscaler repository or Hugging Face and place in the folders required by the custom node.

Recommended Settings

Sampler: Euler or Euler ancestral

Scheduler: Simple, Normal or Beta

Steps: 20–40 CFG Scale: 1.0

Flux Guidance: 3–6

Shift: 4–7

Resolution: freely selectable via the selectLatentSizePlus node (many golden-ratio-friendly presets included, swap orientation with one click)

How to use the workflow

1. Load the JSON in ComfyUI.

2. Select your desired aspect ratio and resolution in the selectLatentSizePlus node (enable swap_orientation if needed).

3. Enter your prompt in the Positive Prompt node (the NdSuperLoraLoader will automatically pull trigger words from your LoRAs).

4. (Optional) Load one or more Flux-compatible LoRAs into the NdSuperLoraLoader – strengths and triggers are handled elegantly.

5. Generate. The first stage produces a high-quality base image.

6. (Optional) Run the SEEDVR2 upscaler subgraph for beautiful 2–4× upscales or even video. It works great on still images too.

7. The SaveImageExtended node gives you full control over filenames, metadata and folder structure.

You can bypass the entire upscaler group if you just want fast test generations. The workflow also works with regular .safetensors Flux checkpoints – simply replace the UnetLoaderGGUF and DualCLIPLoader with the standard nodes.

This is a clean, powerful and future-proof base for all your Flux.1 Dev work – whether you want fast single-LoRA portraits, complex multi-LoRA scenes or high-end upscaled output. I’ve been using it daily on my 6 GB card and it feels snappy and reliable.

Huge thanks to WikkedAI for the original WikkedZITv4 foundation and to the creators of CacheDiT, SageAttention, NdSuperLoraLoader, selectLatentSizePlus and SeedVR2 for making low-VRAM Flux so enjoyable.

If you have questions, feedback or want to share your results – I’m always happy to hear from you in the comments. Enjoy the workflow and happy creating!